A hiker is lost in the forest, and winter weather is setting in. The search needs to be speedy, thorough, and efficient. Searchers send out an autonomous robotic drone. But how does it know where it should go?
New research from Rohan Ghuge , assistant professor of information, risk, and operations management at McCombs School of Business at The University of Texas at Austin, can help.
Unmanned vehicles are becoming a huge industry. The market in 2025 was $29.3 billion and is anticipated to grow to $67.6 billion by 2033, according to market research company Grand View Research. Uses are as varied as military surveillance, package delivery, scientific research, and commercial agriculture.
One common challenge across applications is plotting a course for a vehicle. It might begin with general parameters for the location it needs to search, but parameters will change as it gains more information.
Ghuge wondered how much information was enough. He asks, “How much data do you need to get to say, ‘OK, now I need to stop. I need to change my mind. I need to change the path I’ve been going on.’”
When a searcher doesn’t know where an object or other data are located, there are two ends of the path-finding spectrum, Ghuge says.
In a fully adaptive model, the vehicle is sent on a specified path. It then changes course with every piece of information it gathers, such as photos of a section of the forest.
At the other extreme, a nonadaptive model sets a path and doesn’t divert from it, no matter what new information arises.
Both models have advantages and drawbacks. The fully adaptive model can produce better routes, but continually recomputing the path takes more resources and time. The nonadaptive model might be quicker, but because it’s not adapting to new information, it might not be as effective in achieving its goal, such as locating the lost hiker.
With unmanned vehicles, logistical concerns come into play as well, such as battery life and how often to upload data.
With Rayen Tan and Viswanath Nagarajan of the University of Michigan, Ghuge set out to calculate the optimal balance between accuracy and efficiency.
The researchers tested a hybrid model between nonadaptive and fully adaptive. It sent the robot vehicle on a designated course. Only at the end of each course — what they called “rounds” — did it recompute its path.
The first round was nonadaptive, while subsequent rounds became increasingly adaptive. Running computerized simulations, the team found:
“You don’t really need all the data to make good decisions,” Ghuge says. “If you’re doing the searches sequentially, two or three rounds are sufficient.”
For a business or organization trying to solve a problem quickly — from a search-and-rescue operation to a utility company trying to locate the source of an outage — trading a faster response for a modest cost increase could make a big difference.
“Constantly replanning can itself be very expensive,” Ghuge says. “When you have a robot which changes its solution just a few times, you still get most of the benefit of being fully adaptive. It’s more practical, with a small added cost.”
“ Informative Path Planning With Limited Adaptivity ” is published in INFORMS Journal on Computing .
INFORMS Journal on Computing
Informative Path Planning with Limited Adaptivity
27-May-2026